Dynamic lot sizing for a warm/cold process: Heuristics and insights

(1)

Dynamic lot sizing for a warm/cold process: Heuristics and insights

Ayhan O

¨ zg ¨ur Toy

a,n

, Emre Berk

b

a

Faculty of Industrial Engineering Department, Istanbul Bilgi University, 34060 Ey¨up, Istanbul, Turkey

b

Faculty of Business Administration, Bilkent University, 06800 Bilkent, Ankara, Turkey

a r t i c l e

i n f o

Article history:

Received 15 February 2012 Accepted 10 September 2012 Available online 23 September 2012 Keywords: Lot sizing Warm/cold process Rolling horizon Heuristics

a b s t r a c t

We consider the dynamic lot sizing problem for a warm/cold process where the process can be kept warm at a unit variable cost for the next period if more than a prespeciﬁed quantity has been produced. Exploiting the optimal production plan structures, we develop nine rule-based forward solution heuristics. Proposed heuristics are modiﬁed counterparts of the heuristics developed previously for the classical dynamic lot sizing problem. In a numerical study, we investigate the performance of the proposed heuristics and identify operating environment characteristics where each particular heuristic is the best or among the best. Moreover, for a warm/cold process setting, our numerical studies indicate that, when used on a rolling horizon basis, a heuristic may also perform better costwise than a solution obtained using a dynamic programming approach.

1. Introduction

In this paper, we consider the problem of dynamic lot sizing for a special type of production processes. The dynamic lot sizing problem is deﬁned as the determination of the production plan which minimizes the total (ﬁxed setup, holding and variable production) costs incurred over the planning horizon for a storable item facing known demands.

Recently, the notion of a ‘‘warm/cold process’’ has been introduced into the scheduling literature (Toy and Berk, 2006). A warm/cold process is deﬁned as a production process that can be kept warm for the next period if a minimum amount (the so-called warm threshold) has been produced in the current period and would be cold, otherwise. Production environments where the physical nature of the production technology dictates that the processes be literally kept warm in certain periods to avoid expensive shutdown/startups are typical in glass, steel and ceramic production. Robinson and Sahin (2001) provide other examples in food and petrochemical industries where certain cleanup and inspection operations can be avoided in the next period if the quantity produced in the current period exceeds a certain threshold. Production processes where production rates can be varied also fall into the warm/cold process category. The upper bound on the production rate is the physical capacity of the production process and the lower bound corresponds to the warm threshold, below which the process cannot be kept running into the next period without incurring a setup. Such variable

production rates can be found in both discrete item manufactur-ing and process industries. Change in production rate can be obtained at either zero or positive cost depending on the char-acteristics of the employed technology. The additional variable cost is, then, the variable cost of keeping the process warm onto the next period.

As the above examples illustrate, the dynamic lot-sizing problem in the presence of production quantity—dependent warm/cold processes is a common problem. This problem, in the presence of no shortages, has been formulated and solved optimally byToy and Berk (2006)using a dynamic programming approach with an OðN3

Þforward algorithm where N denotes the problem horizon length. Later, they extend their results to the case where some of the demands may be lost under a proﬁt maximization objective (Berk et al., 2008).

The dynamic lot sizing problem for a warm/cold process is a generalization of the so-called classical problem which was ﬁrst analyzed by Wagner and Whitin (1958). The classical problem assumes uncapacitated production and no shortages.Wagner and Whitin (1958) provide a dynamic programming solution algo-rithm and structural results on the optimal solution. Their fundamental contribution lies in establishing the existence of planning horizons, which makes forward solution algorithms possible. Although the optimal solution structure is known, the complexity of obtaining it (shown to be OðN log NÞ in general by

Federgruen and Tzur, 1991;Wagelmans et al., 1992;Feng et al., 2011for constant capacities) has stimulated a stream of research that focuses on developing lot sizing heuristics based on simple stopping rules, such as Silver–Meal (Silver and Meal, 1973), Part-Period Balancing (DeMatteis, 1968), Least Unit Cost, Economic Order Interval, McLauren’s Order Moment (Vollmann et al., 1997), Least Total Cost (Narasimhan and McLeavy, 1995), Groff’s Contents lists available atSciVerse ScienceDirect

journal homepage:www.elsevier.com/locate/ijpe

Int. J. Production Economics

n

Corresponding author.

E-mail addresses: ozgur.toy@bilgi.edu.tr (A.O¨ . Toy), eberk@bilkent.edu.tr (E. Berk).

(2)

Algorithm (Groff, 1979). (See alsoSahin et al., 2008; Narayanan and Robinson, 2010.)

Further results on the lot sizing problem are found in the literature on its extension to the capacitated production settings. The capacitated lot sizing problem (CLSP) is related to the lot sizing problem for a warm/cold process under certain conditions (see Toy and Berk, 2006). The CLSP has been ﬁrst studied by

Manne (1958)and has been shown to be NP-hard byFlorian et al. (1980). Reviews of the works on CLSP (along with the uncapaci-tated versions) are byBrahimi et al. (2006)andQuadt and Kuhn (2008), who include extensions of the problem, and Buschk ¨uhl et al. (2010). Recent analytical studies have focused on novel solution approaches.Heuvel and Wagelmans (2006)develop an OðT2

Þ algorithm. Pochet and Wolsey (2010) provide a mixed integer programming reformulation that can be solved with LP-relaxation to optimality under reasonable conditions.Chubanov et al. (2008)andNg et al. (2010)introduce polynomial approx-imations.Hardin et al. (2007) analyze the quality of bounds by fast algorithms. Reviews of meta-heuristic approaches to the CLSP can be found inStaggemeier and Clark (2001),Jans and Degraeve (2007)and inGuner Goren et al. (2010)on genetic algorithms for lot sizing. A recent review of related works appears also inGlock (2010).

Rule-based heuristics in rolling horizon environments have been studied byStadtler (2000),Simpson (2001), andHeuvel and Wagelmans (2005). The work herein joins this stream by con-sidering the dynamic lot sizing problem for a warm/cold process. Specifically, we propose rule-based lot sizing heuristics for the problem and examine the efficacy of such rules. To the best of our knowledge, this is the first work that studies lot sizing rules for the operating environment where the production process can be kept warm at some cost if production quantity in a period exceeds a threshold value. We believe that our contributions lie in developing a number of heuristics which perform well in certain operational environments and in identifying such regions for selecting a particular heuristic. We consider the application of the proposed heuristics in a static setting as well as on a rolling horizon basis as it is the practice. The available commercial ERP software (e.g., SAP) still offer well-known heuristics for the classical lot sizing problem as options for decision-makers along with the ‘optimal’ solution algorithms in their manufacturing modules. For the conventional production environments, the benefits of heuristics include the ease of use, smoother produc-tion schedules and more intuiproduc-tion for the trade-offs. Moreover, for a warm/cold process setting, our numerical studies indicate that, when used on a rolling horizon basis, a heuristic may also perform better costwise than a solution obtained using a dynamic pro-gramming approach. This finding is consistent with similar studies on the classical problem (Stadtler, 2000; Heuvel and Wagelmans, 2005). Hence, investigation of heuristics for warm/ cold process settings may be financially beneficial in practice as well as from a purely theoretical perspective. Our work extends the heuristics literature on the dynamic lot sizing problem.

The rest of the paper is organized as follows: InSection 2, we introduce the basic assumptions of our model, formulate the optimization problem and present some key results. InSection 3, we present some theoretical results on an economic production quantity (EPQ) model that we use as a continuous counterpart of a warm/cold process to develop some of our heuristics. InSection 4, we introduce and construct nine lot sizing heuristics for a warm/ cold process. InSection 5, we present a numerical study and discuss our ﬁndings in regards to the cost performance of the proposed heuristics. In our numerical study, we provide results on the performance distribution of individual heuristics, on the rankings of the heuristics, on identifying the operating environment where a particular heuristic may perform best and on the impact of

planning horizon lengths when production plans are made and executed on a rolling horizon basis.

2. Model: assumptions and formulation

We consider the operational setting in Toy and Berk (2006)

with time-invariant system and cost parameters. We assume that the length of the problem horizon, N is ﬁnite and known. Demand in period t, denoted by Dt ðt ¼ 1,2, . . . ,NÞ, is non-negative and known, but may be different over the problem horizon. No shortages are allowed; that is, the amount demanded in a period has to be produced in or before its period. The amount of production in period t is denoted by xt. If xt40, the production

indicator

d

tis 1, zero otherwise. The inventory on hand at the end

of period t is denoted by yt ð ¼yt1þxtDtÞ. Inventory holding

cost per unit of ending inventory is h per period. Without loss of generality, we assume that the initial inventory level is zero. We assume that unit production cost is c but may be omitted in the analysis since all demands must be met over the horizon.

Production quantity in a period cannot exceed the capacity, R. For feasibility, we assume that, for any t, there exists a j(t) for whichPjðtÞ_{i ¼ t}DirðjðtÞt þ1ÞR for t rjðtÞrN, 1rt rN. This

condi-tion guarantees that any subset of demands can be produced within the horizon; a special case of the condition is satisﬁed when DtrR for all t. We consider a warm/cold production

process: The production process may be kept warm onto the beginning of period t if the production quantity in the previous period is at or above a threshold value Q ; that is, xt1ZQ . Otherwise, the process cannot be kept warm and is cold. Let zt indicate the warm/cold status of the process as period t starts; it attains a value of 0 if the process is warm and 1, otherwise. In order to keep the process warm onto period t, warming cost

o

is charged for every unit of unused capacity in period t1. That is, the warming cost incurred in period t1 would be

o

ðRxt1Þ

monetary units. Note that, even if the quantity produced in period t1 is at least Q , it may not be optimal to keep the process warm onto the next period if there would not be any production during the next period. In such instances, there will be no warming costs incurred although xt1ZQ since xt¼0. We assume that a warm

process requires no setup (and, hence incurs no setup cost) but a cold process requires a cold setup with a ﬁxed cost K ð 40Þ if production is to be done in the period. Finally, we assume that h4

o

which ensures the Wagner–Whitin type cost structure, and that the warm/cold process threshold is between the point of indifference and the capacity, RðK=

o

ÞoQ rR. (For the implica-tions of these assumpimplica-tions, seeToy and Berk, 2006.)

The objective is to ﬁnd a production plan xtZ0 ðt ¼ 1,2, . . . ,NÞ (timing and amount of production), such that all demands are met at minimum total cost over the horizon. Let X ¼ fx1, . . . ,xNg

denote a feasible production plan constructed over periods 1 through N;

G

t be the variable cost incurred within period t

computed as

G

t¼hytþ

o

ðRxtÞ

d

t þ 1ð1zt þ 1Þ under the given

production plan; and THC denote the total horizon cost. Then, the optimization problem (P) can be formally stated as follows: min

X THC ¼

XN t ¼ 1

ðK

d

tztþ

G

tÞ

subject to 0rxtrR, yt1þxtZD_t, for all t; z₁¼1,

d

N þ 1¼0,

zN þ 1¼1, and y0¼yN¼0.

Let Lu,vrepresent a subset of X between periods u and v1

(inclusive) such that the starting inventory in period u and ending inventory in period v1 are zero and production is done in all periods u through m to cover the demands for periods u through v1. Formally, Lu,v¼ fxt9xt40,t ¼ u, . . . ,m; xt¼0 for t ¼ m þ 1, . . . ,v1;

(3)

terminology, we shall refer to Lu,v as a production lot (for period u

through v1), the total quantity produced in the production lot as its lot size and ðvuÞ as the production lot run length. Under the assumed cost structure, we have the following result:

Proposition 1 (Optimal Schedule within a Production Lot). The optimal (total cost minimizing) structure of a production lot Lu,v is

as follows: (i) xt¼maxðQ ,½Dtyt1Þ, for urt rm1 and mu40,

(ii) xm¼

e

, where

e

¼ ½Pv1i ¼ mDiym1þoR and muZ0, and

(iii) xt¼0 for t ¼ m þ1, . . . ,v1.

The proof rests on showing that any other production plan for the periods u through v1 with the given zero inventory constraints would result in higher total costs and hence cannot be optimal. Note that, typically, in the uncapacitated, classical setting, each production lot corresponds to a single period of production followed by non-production periods (with u ¼m); only in rare cases where setup costs are (comparably) very small, one would get a lot-for-lot schedule. But, for a warm/cold process, a production may contain a number of periods in which production is done in succession (with urm). Also note that the optimal production plan for problem (P) consists of production lots that have the same structure as that given above (see Theorems 1–4 in

Toy and Berk, 2006).

Similarly, in devising and implementing the forward solution heuristics below, we will use this key result. All production lots Lu,vwill be assumed to have the above optimal structure. Hence,

heuristics will result in lot sizes (i.e., the selection of some period u and v) that may be sub-optimal but the structure of the production plan will conform to the optimal solution structure.

Let Cu,v be the variable cost incurred over periods u through

v1 by the production lot Lu,v,

Cu,v¼ X v1 t ¼ u

G

t¼ X v1 t ¼ u hytþ X m1 t ¼ u

o

ðRxtÞ

In the following Lemma, we state how incremental updating of production plans and costs is performed as the production lot length is extended by one period, i.e., how the production lot Lu,v þ 1¼ fx0t9x0t40, t ¼ u, . . . ,m0; x0t¼0 for t ¼ m0þ1, . . . ,v1g and

the corresponding cost Cu,v þ 1 are obtained dynamically from

Lu,v¼ fxt9xt40,t ¼ u, . . . ,m; xt¼0 for t ¼ m þ1, . . . ,v1g and Cu,v,

respectively.

Lemma 1. Let Lu,v¼ fxt9xt40,t ¼ u, . . . ,m; xt¼0 for t ¼ m þ 1, . . . ,

v1g, y0

t¼y0t1þx0tDt and Et¼ ðmaxðQ ,½Dty0t1 þ_Þðh

_o

Þ þ R

o

Þ=h.

(a) Lu,v þ 1¼ fx0t9x0t40,t ¼ u, . . . ,m0; x0t¼0 for t ¼ m0þ1, . . . ,v1g

where x0 t¼ xt for urt om Dvþxm for t ¼ m if DvþxmrEt maxðQ ,½Dty0t1 þ_Þ _{for t ¼ m if D} vþxm4Et Dvþxm Xt1 j ¼ m x0 j 2 4 3 5 þ for mot ovþ1 if Dvþxm Xt1 j ¼ m x0 j 2 4 3 5 þ rEt maxðQ ,½Dty0t1 þ Þ for mot ovþ1 if Dvþxm Xt1 j ¼ m x0 j 2 4 3 5 þ 4Et 8 > > > > > > > > > > > > > > > > > > > > > > > > > < > > > > > > > > > > > > > > > > > > > > > > > > > : and m0_¼_{maxft : x}0 t40,urt ovþ1g; (b) Cu,v þ 1¼Cu,vþPv1i ¼ mhð Pi j ¼ mx0jxmÞ þPm01i ¼ m

o

ðRx0iÞ.

Proof. (a) Immediately follows from the optimal schedule structure. (b) Let Cu,v be the total variable cost associated with the

production lot Luvas deﬁned before. Hence, Cu,v þ 1Cu,v¼ Xv i ¼ u hX i j ¼ u ðx0 jDjÞ þ Xv i ¼ u Iðx0 i40Þðx 0 i þ 140Þ

o

ðRx 0 iÞ 2 4 3 5 X v1 i ¼ u hX i j ¼ u ðxjDjÞ þ X m1 i ¼ u

o

ðRxiÞ 2 4 3 5 ¼ X m1 i ¼ u hX i j ¼ u ðx0 jDjÞ þh Xm j ¼ u ðx0 jDjÞ 2 4 þ X v1 i ¼ m þ 1 hX i j ¼ u ðx0 jDjÞ þh Xv j ¼ u ðx0 jDjÞ þX m1 i ¼ u

o

ðRx0 iÞ þ Xv i ¼ m Iðx0 i40Þðx 0 i þ 140Þ

o

ðRx 0 iÞ # X m1 i ¼ u hX i j ¼ u ðxjDjÞ þh Xm j ¼ u ðxjDjÞ 2 4 þ X v1 i ¼ m þ 1 hX i j ¼ u ðxjDjÞ þ X m1 i ¼ u

o

ðRxiÞ 3 5 ¼ X v i ¼ m h X i j ¼ m x0 jxm 0 @ 1 AþX v i ¼ m Iðx0 i40Þðx 0 i þ 140Þ

o

ðRx 0 iÞ

which reduces to the result. &

Before introducing the heuristics, we will explore some prop-erties of the continuous review economic production quantity (EPQ) counterpart of a warm/cold process in the next section.

3. Economic production quantity for a warm/cold process In this section, we construct the continuous review economic production quantity (EPQ) counterpart of a warm/cold process. The model derived herein serves as the foundation of the EPQ-based heuristics to be discussed later. The model assumptions are as follows: Demand rate is deterministic and constant, d. There is a constant production rate, p which is a decision variable and may take on values over ½maxðQ ,dÞ,R per unit time where Q denotes the physical threshold for a warm process. The cost rate of operating with a particular value of the production rate p0 is given by

o

ðRp0Þ. This cost is analogous to the warming cost in

the sense that it keeps the production process running. We omit the unit production cost due to material usage, etc. as in the periodic review problem at hand. Hence, the effective production cost rate is

o

ðRp0Þ. A cycle is deﬁned as the time between two

consecutive instances of a cold setup, which initiates a production run. Each cycle consists of two phases – a production and a non-production phase – as in the classical EPQ setting. During the production phase, q units are produced and inventory is accumu-lated at a rate that is in excess of the demand rate; during the non-production phase, the accumulated inventory is used up to satisfy the demand until the next production run. Each unit of inventory held per unit time incurs a carrying cost h. The objective is to ﬁnd the production quantity and the produc-tion rate that minimize the total cost per unit time.

The cycle length for a given production quantity q and a production rate p is given by

(4)

The length of the production phase in a cycle is q/p and the maximum inventory level at the end of the production phase is ðpdÞðq=pÞ. In each cycle, there is a single cold setup by deﬁnition, and warming cost is incurred during the entire production phase. Therefore, the total (setup, holding and production/warming) cost per cycle is CCðq,pÞ ¼ K þq 2_ðpdÞ 2pd h þ qðRpÞ p

o

The total cost rate is TCðq,pÞ ¼ CCðq,pÞ=CLðq,pÞ ¼ Kd=q þ ðqðpdÞ=2pÞhþ ½

o

ðRpÞ=pd. Note that the cost rate expression is similar in structure to that of an economic production quantity model in the presence of (i) production rate dependent unit production costs or (ii) cost of selecting a production rate (e.g. Khouja and Mehrez, 1994;Larsen, 1997).The unique optimizer ^qðpÞ of TCðq,pÞ for a given value of production rate p is given by the classical EPQ formula:

^

qðpÞ ¼pffiffiffiffiffiffiffiffiffiffiffiffiffiffi2Kd=hpffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffip=ðpdÞ with the corresponding optimal cost rate TC ðpÞ ¼c pffiffiffiffiffiffiffiffiffiffiffi2KdhpffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiðpdÞ=pþd

o

ðRpÞ=p. (Note, when p¼ d, the process is run continuously ð ^qðpÞ_{-1Þ resulting in} c

TC ðp ¼ dÞ ¼

o

ðRdÞ.) The unique extremum of cTC ðpÞ is given by ~

p ¼ 2

o

2_R2

d=ð2

o

2_R2

dKhÞ. However, it is not possible to say that ~p is a global minimum (or maximum) since the principal minors of the Hessian are of mixed signs. Furthermore, it may not lie within the feasible region. Therefore, to obtain the optimal production rate pn

, we need to check the total cost rates attained when the production rate takes on the extremum value (if feasible) and its two boundaries; pn

¼arg min_{p A fmaxðQ ,dÞ,Rg[ ^}_PTC ðpÞ where we deﬁne the set ^c P ¼ fr : r ¼ ~p; maxðQ ,dÞrr rRg which is non-empty only for feasible ~p. Then, we have dTCn

¼ cTC ðpn

Þand cqn_{¼ ^}_qðpn

Þ; we also let dCCn denote the corresponding cycle cost.

4. Heuristics

In this section, we develop the proposed lot sizing heuristics. Solving the problem (P) over a problem horizon from period i0

to period N0

by a forward heuristic means that a production plan is obtained in a forward manner (starting from period i0

and proceeding ahead up to period N0_{) by employing a pre-speciﬁed stopping rule to}

determine the production lots over the horizon. The stopping rule of a heuristic dictates when a production lot starting in period u should terminate; at the termination of one production lot, it is assumed that another starts. Due to Proposition 1, the structure of the production lot is known. Applying the rule over the horizon to generate successive production lots, one obtains the production plan X under the heuristic. The total horizon cost THC under the heuristic is then computed using the production plan X. The implementation of the forward heuristics herein follows the pseudo-code below.

Program_FindProductionPlanði0 ,N0 Þ begin i :¼ i0 while ðirN0 Þ u :¼ i; n :¼ 1; stoping_rule :¼false generate Lu,u þ n

while (stopping_rule¼ ¼false) generate Lu,u þ n þ 1

if stopping condition is satisﬁed then do stopping_rule :¼true

compute ^n (number of periods in the production lot)

i :¼ u þ ^n else n :¼ n þ1 end

We propose nine heuristics for the dynamic lot sizing problem for a warm/cold process. They are adaptations or modiﬁcations for the warm/cold process setting of the heuristics that have been developed for the classical, uncapacitated lot sizing problem. As we show in our numerical analysis, some of the proposed heuristics do not perform as well for certain system and cost parameters and demand patterns. However, there is not one particular heuristic that performs best in all experiments either. Below, we discuss the construction of our heuristics and state the stopping rules for each heuristic.

Heuristic #1 is developed in the same essence as the Silver– Meal heuristic in the classical, uncapacitated setting (Silver and Meal, 1973); it is an adaptation of the Silver–Meal heuristic for a warm/cold process. It rests on the comparison of the total cost per period. A production lot starting at u terminates at period u þ ^n1 after which the total cost rate increases for the ﬁrst time. Formally, the stopping rule for this heuristic can be stated as K þCu,u þ n=noK þCu,u þ n þ 1=ðn þ 1Þ with n ¼ 1þ maxfn : K þ^

Cu,u þ n=n ZK þ Cu,u þ n þ 1=ðn þ 1Þg.

To illustrate the mechanics of the proposed heuristics, we shall use an illustrative example with the following parameter set: K¼10, h¼ 1,

o

¼0:95, R¼8, and Q¼5 with N¼ 15 and demands over the horizon, D ¼ f6,4,2,2,4,7,5,6,4,8,4,5,8,1,4g. The optimal production plan for periods 1 through 15 for this example is found as ff6,5,3,0g, f5,6,5,6,5,7,5,5,7,5,0gg with a corresponding total cost of 55.70. Note that the production plan is presented as a sequence of produc-tion lots and each producproduc-tion lot is deﬁned as a sequence of the production done on each consecutive period. For Heuristic #1, starting with the ﬁrst period, we construct the production lot Lu,u þ n, compute the stopping criterion

r

u,u þ n¼ ðK þCu,u þ nÞ=n and

apply the rule as follows. L12¼ f6g,

r

12¼ ð10 þ0Þ=ð21Þ ¼ 10;

L13¼ f6,4g,

r

13¼ ð10 þ 0:95ð86ÞÞ=ð31Þ ¼ 5:95 ðo

r

12Þ; L14¼

f6,6,0g,

r

14¼ ð10 þ0:95ð86Þ þ1ð0 þ 64ÞÞ=ð41Þ ¼ 4:63 ðo

r

13Þ;

L15¼ f6,5,3,0g,

r

15¼ ð10 þ ½0:95ð86Þ þ 0:95ð85Þ þ½1ð0 þ 54Þ þ

1ð1 þ 32ÞÞ=ð51Þ ¼ 4:44ðo

r

14Þ; L16¼ f6,5,7,0,0g,

r

16¼5:15

ð4

r

15Þ. We freeze the production plan for periods 1 through 4,

f6,5,3,0g, and start from period 5. Proceeding in a similar fashion, we get the production plan for periods 1 through 15 as ff6,5,3,0g, f5,6,5,6,5,7,4g, f5,8,5,0gg with a corresponding total cost of 59.90.

Heuristic #2 is similar in construction to the Part Period Balancing (PPB) heuristic in the classical, uncapacitated setting (DeMatteis, 1968); it is an adaptation of the PPB heuristic for a warm/cold process. It rests on the comparison of the cost of a production lot against a cold setup. A production lot starting at u terminates at period u þ ^n1 after which the variable cost exceeds the ﬁxed cold setup cost for the ﬁrst time. Formally, the stopping rule can be stated as Cu,u þ n4K with

^

n ¼ maxfn : Cu,u þ nrKg.

Consider the same illustrative setting given above. Starting with the ﬁrst period, we construct the production lot Lu,u þ n,

compute the stopping criterion

r

un¼Cu,u þ n and apply the rule

as follows. L12¼ f6g,

r

12¼0; L13¼ f6,4g,

r

13¼0:95ð86Þ ¼ 1:90

ðoK ¼ 10Þ; L14¼ f6,6,0g,

r

14¼0:95ð86Þ þ 1ð0 þ 64Þ ¼ 3:90

ðoKÞ; L15¼ f6,5,3,0g,

r

15¼ ½0:95ð86Þ þ 0:95ð85Þ þ ½1ð0 þ

54Þ þ 1ð1 þ32Þ ¼ 7:75ðoKÞ; L16¼ f6,5,7,0,0g,

r

16¼15:75

ð4KÞ. We freeze the production plan for periods 1 through 4, f6,5,3,0g, and start from period 5. Proceeding in a similar fashion, we get the production plan for periods 1 through 15 as ff6,5,3,0g, f5,6,5,6g, f5,7,5,4g, f8,5,0gg with a corresponding total cost of 60.90.

Heuristic #3 is the adaptation for a warm/cold process of the Least Unit Cost heuristic (LUC) in the classical, uncapacitated setting (Vollmann et al., 1997). It rests on the total cost per unit comparison. A production lot starting at u terminates at period u þ ^n1 after which the total cost rate increases for the ﬁrst time. Formally,

(5)

the stopping rule can be stated as ðK þCu,u þ nÞ=Pu þ n1i ¼ u DioðK þ

Cu,u þ n þ 1Þ=Pu þ ni ¼ uDi with ^n ¼ 1 þ maxfn : ðK þ Cu,u þ nÞ=Pu þ n1i ¼ u DiZ ðK þCu,u þ n þ 1Þ=Pu þ ni ¼ uDig.

For Heuristic #3, starting with the ﬁrst period, we construct the production lot Lu,u þ n, compute the stopping criterion

r

un¼ ðK þ Cu,u þ nÞ=Pu þ n1i ¼ u Di and apply the rule as follows.

L12¼ f6g,

r

12¼ ð10 þ0Þ=ð6Þ ¼ 1:67; L13¼ f6,4g,

r

13¼ ð10 þ 0:95

ð86ÞÞ=ð6 þ 4Þ ¼ 1:19 ðo

r

12Þ; L14¼ f6,6,0g,

r

14¼ ð10 þ 0:95ð8

6Þ þ 1ð0 þ 64ÞÞ=ð6 þ 4 þ2Þ ¼ 1:16 ðo

r

13Þ; L15¼ f6,5,3,0g,

r

15¼

ð10 þ ½0:95ð86Þ þ 0:95ð85Þ þ ½1ð0 þ54Þ þ1ð1 þ 32ÞÞ=ð6 þ4 þ2þ 2Þ ¼ 1:27 ð 4

r

14Þ. We freeze the production plan for periods

1 through 3, f6,6,0g, and start from period 4. Proceeding in a similar fashion, we get the production plan for periods 1 through 15 as ff6,6,0g, f5,5,5,5,5,5,6,5,5,7,1g,f4gg with a corresponding total cost of 75.55.

Heuristic #4 is the adaptation for a warm/cold process of the Least Total Cost (LTC) heuristic in the classical, uncapacitated setting (Narasimhan and McLeavy, 1995). A production lot start-ing at u terminates at period u þ ^n1 at which the absolute difference between the ﬁxed cold setup cost and the total variable cost is minimum. Formally, the stopping rule can be stated as Cu,u þ n4K with ^n ¼ arg minn A f ~n, ~n þ 1g9KCu,u þ n9 where ~n ¼ max

fi : Cu,u þ irKg.

Starting with the ﬁrst period, we construct the production lot Lu,u þ n, compute the stopping criterion

r

un¼Cu,u þ n and apply the

rule as follows. L12¼ f6g,

r

12¼0; L13¼ f6,4g,

r

13¼0:95 ð86Þ ¼

1:90 ðoK ¼ 10Þ; L14¼ f6,6,0g,

r

14¼ ½0:95ð86Þ þ1ð0 þ 64Þ ¼ 3:90

ðo10Þ; L15¼ f6,5,3,0g,

r

15¼ ½0:95ð86Þ þ 0:95ð85Þ þ ½1ð0 þ 54Þ þ

1ð1 þ32Þ ¼ 7:80 ðo10Þ; L16¼ f6,5,7,0,0g,

r

16¼15:75 ð 410Þ.

Note that the difference for period 5 is larger than that for period 4; minð91015:759,9107:809Þ ¼ 2:20. Hence, we freeze the produc-tion plan for periods 1 through 4, f6,5,3,0g, and start from period 5. Proceeding in a similar fashion, we get the production plan for periods 1 through 15 as ff6,5,3,0g, f5,6,5,6,4g,f8,5,5,7,1g,f4gg with a corresponding total cost of 66.95.

For the classical, uncapacitated setting, a group of heuristics have been proposed in the literature based on a continuous review (economic order quantity, EOQ) approximation of the periodic review production environment: Groff’s heuristic (Groff, 1979), the economic order interval (EOI) heuristic, and McLauren’s Order Moment (MOM) heuristic (seeVollmann et al., 1997). In these heuristics, the demand stream over the problem horizon is approximated by a constant rate demand process with the same mean; and, the cost trade-offs inherent in the lot sizing problem for the periodic setting are approximated by the cost trade-offs based on the EOQ model. This class of heuristics has been shown to be effective in certain operating environments. For the warm/cold production processes, we propose four heur-istics that are adaptations or modifications of these classical setting heuristics. The existence of a warm/cold process implies the capability of keeping the process warm (running) by selecting minimum production quantities (at or above the warm threshold) and incurring a warming cost. This suggests an analogy between the warm/cold setting and a production environment with a finite production rate. Hence, the construction of our heuristics is based on the approximation of the periodic review, warm/cold produc-tion setting via the continuous review, economic producproduc-tion quantity (EPQ) model. It attempts to capture the warm/cold capability through the finite production rate that can be selected at a cost. For these heuristics, we have the following definitions. Let D denote the constant demand rate for the problem horizon; d

TCn

denote the optimal total cost rate and cqn _{denote the optimal} economic production quantity for the equivalent EPQ model. The equivalent EPQ model assumes constant demand rate D and a constant production rate pn _{whose computation involves} the warm threshold Q, the physical production capacity R and

the warming cost

o

. We refer the reader to Section 3 for the

development of the equivalent EPQ approximation and

the derivation of the basic operational entities above ( dTCn , cqn and pn_{) used in the construction of the heuristics. We introduce} these four heuristics (Heuristics #5–8) below.

Heuristic #5 is the adaptation for a warm/cold process of the Economic Order Interval (EOI) heuristic in the classical, uncapa-citated setting (Vollmann et al., 1997). It is based on an equivalent (continuous review) economic production quantity cqn _obtained by assuming a constant demand rate D over the horizon and imposing a constant production rate pn

. By construction, with this rule, a production plan consists of production lots of equal size (except possibly the last lot). In this heuristic, a production lot starting at u terminates at period u þ ^n1 when the production lot reaches a length equal to the largest integer that is smaller than the economic production interval (EPI). The stopping rule can be stated as n 4 cqn_{=D with ^}_{n ¼ bc}_qn_{=Dc with c}_qn

¼ND if pn ¼D. For the given demand stream and cost parameters of the illustrative example, we have D ¼ 4:67, ~p ¼ 7:83, cTC ðQ Þ ¼ 5:15;

c

TC ð ~pÞ ¼ 6:237; cTC ðRÞ ¼ 6:236; cTCn¼5:15; pn

¼5:00 and cqn_¼_37:42 for the heuristics based on the EPQ model. For Heuristic #5,

^

n ¼ b37:42=4:67c ¼ 8 which gives the production plan for periods 1 through 15 as ff6,5,5,5,5,5,5,0g, f5,7,5,5,7,5,0gg with a corre-sponding total cost of 85.60.

In the classical, uncapacitated setting,Groff (1979)proposes a heuristic that uses an EOQ approximation to the dynamic lot sizing problem. It is based on the comparison of marginal changes in the ﬁxed and variable cost components when the length of a production lot is extended by one period. We develop Heuristic #6 in a similar fashion using, instead, an EPQ approximation for a warm/cold process. Consider the production lot Lu,u þ n. Let ~Cu,u þ n

denote the total holding and warming costs within the production lot as approximated by its EPQ counterpart for a constant demand rate D over the horizon and imposing a constant production rate pn

; and, let ½K þ ~Cu,u þ nðN=nÞ denote the corresponding

approx-imate total cost over the horizon as in Groff. We obtain ~Cu,u þ nas

follows. With constant demand rate over the horizon, D and the imposed production rate pn

, the maximum inventory level within the production lot is Imax¼Pu þ n1i ¼ u DiðpnDÞ=pn. The cost of the

inventory carried within the lot is approximated as hnImax=2.

Noting thatPu þ n1i ¼ u Di=pngives the length of the production run

within the lot, the corresponding warming cost within the lot is computed as ðRpn

Þ

o

Pu þ n1i ¼ u Di=pn, which we approximate as

ðRpn Þ

o

nD=pn

by replacing the total demand by its average. Then, ~ Cu,u þ n¼hn X u þ n1 i ¼ u Diðp n DÞ=ð2pn Þ þ ðRpn Þ

o

nD=pn

Under Heuristic #6, a production lot starting at u terminates at period u þ ^n1 after which the approximate total cost over the horizon increases for the ﬁrst time. Formally, the stopping rule can be stated as ½K þ ~Cu,u þ nðN=nÞo½K þ ~Cu,u þ n þ 1ðN=ðn þ1ÞÞ which

simpliﬁes to nðn þ 1ÞDu þ n4 ½cqn2=D ð ¼2Kp

n =hðpn

DÞÞ with

^

n ¼ 1þ max fn : nðn þ 1ÞDu þ no½cqn2=Dg with cqn¼ND if pn¼D.

Starting with the ﬁrst period, we construct the production lot Lu,u þ n, compute the stopping criterion

r

un¼ ½K þ ~Cu,u þ nðN=nÞ and

apply the rule as follows. L12¼ f6g,

r

12¼192:9; L13¼ f6,4g,

r

13¼119:9 ðo

r

12Þ; L14¼ f6,6,0g,

r

14¼95:9 ðo

r

13Þ; L15¼

f6,5,3,0g,

r

15¼84:4 ðo

r

14Þ; L16¼ f6,5,7,0,0g,

r

16¼78:9 ðo

r

15Þ;

L17¼ f6,5,5,5,4,0g,

r

17¼77:4 ðo

r

16Þ; L18¼ f6,5,5,5,5,4,0g,

r

18¼76:33 ðo

r

17Þ; L19¼ f6,5,5,5,5,5,4,0g,

r

19¼76:65 ð 4

r

18Þ.

We freeze the production plan for periods 1 through 8, f6,5,5,5,5,4,0g, and start from period 9. Proceeding in a similar fashion, we get the production plan for periods 1 through 15 as ff6,5,5,5,5,4,0g, f6,5,7,5,5,7,5,0gg with a corresponding total cost of 77.65.

(6)

Next, we develop the two heuristics (Heuristics #7 and #8) that are modiﬁcations of the McLauren’s Order Moment (MOM) heur-istic in the classical, uncapacitated setting (Vollmann et al., 1997). The construction logic of the MOM heuristic is similar to that of the PPB heuristic; it determines the lot size for individual orders by matching the number of accumulated part periods to the number that would be incurred if an order of size equal to EOQ were placed under conditions of constant demand. The construction of Heuristics #7 and #8 is similar; an EPQ equivalent is used with constant demand rate over the horizon, D and the imposed production rate pn_{instead of the EOQ. In Heuristic #7, we use the total cost within} the lot as the basis for the stopping rule. A production lot starting at u terminates at period u þ ^n1 after which the total cost within the lot exceeds the benchmark value (cost for an EPQ cycle). It can be formally stated as K þCu,u þ n4 dCC

n

with ^n ¼ maxfn : K þ Cu,u þ nr

d CCn

g. In Heuristic #8, we use the production quantity within the lot as the basis for the stopping rule. A production lot starting at u terminates at period u þ ^n1 after which the total demand to be satisﬁed by the lot exceeds the benchmark value (an EPQ). It can be formally stated as Pu þ n1i ¼ u Di4 cqn with ^n ¼ maxfn :Pu þ n1i ¼ u Dir

c

qn_g_{with c}_qn_¼_{ND if p}n ¼D.

Consider the illustrative example. For Heuristic #7, starting with the ﬁrst period, we construct the production lot Lu,u þ n, compute

the stopping criterion

r

un¼K þ Cu,u þ n and apply the rule as

follows. L12¼ f6g,

r

12¼10; L13¼ f6,4g,

r

13¼10 þ 0:95ð86Þ ¼ 11:90 ðoðcqn_{=DÞ d}_TCn ¼ ðð37:41Þð5:15Þ=4:67Þ ¼ 41:33Þ; L14¼ f6,6,0g, r₁₄¼10 þ0:95ð86Þ þ 1ð0 þ 64Þ ¼ 13:90 ðo41:33Þ; L15¼ f6,5,3,0g,

r

₁₅¼10 þ ½0:95ð86Þ þ 0:95ð85Þþ ½1ð0 þ 54Þ þ 1ð1þ 32Þ ¼ 17:75 ðo41,33Þ; L16¼ f6,5,7,0,0g, r16¼25:75ðo41:33Þ; L17¼ f6,5,5,5,4,0g,

r

17¼39:45 ðo41,33Þ; L18¼ f6,5,5,5,5,4,0g,

r

18¼48:3 ð 4 41,33Þ.

We freeze the production plan for periods 1 through 7, f6,5,5,5,4,0g, and start from period 8. Proceeding in a similar fashion, we get the production plan for periods 1 through 15 as ff6,5,5,5,4,0g, f5,6,5,7,5,5,7,5,0gg with a corresponding total cost of 71.65. For Heuristic #8, starting with the ﬁrst period, we construct the production lot Lu,u þ n, compute the stopping criterion

r

un¼

Pu þ n1

i ¼ u Di and apply the rule as follows. L12¼ f6g, r12¼6

ðocqn_¼_37:42Þ_; _L 13¼ f6,4g, r13¼6 þ 4 ¼ 10 ðo37:42Þ; L14¼ f6,6,0g, r₁₄¼12 ðo14:97Þ; L15¼ f6,5,3,0g, r15¼14 ðo37:42Þ; L16¼ f6,5,7,0,0g, r16¼18 ðo37:42Þ; L17¼ f6,5,5,5,4,0g;r17¼25 ðo37:42Þ; L18¼ f6,5,5,5,5,4,0g,

r

18¼30 ðo37:42Þ;L19¼ f6,5,5,5,5,5,5,0g,

r

19¼36 ðo37:42Þ; L1,10¼ f6,5,5,5,5,5,5,4,0g,

r

1,10¼40 ð 437:42Þ. We

freeze the production plan for periods 1 through 9, f6,5,5,5, 5,5,5,0g, and start from period 10. Proceeding in a similar fashion, we get the production plan for periods 1 through 15 as ff6,5,5,5,5,5,5,0g, f5,7,5,5,7,5,0gg with a corresponding total cost of 85.60.

Finally, we propose Heuristic #9 which is a modiﬁcation for a warm/cold process of the Wagner–Whitin solution algorithm in the classical, uncapacitated setting (Wagner and Whitin, 1958). The construction of the heuristic is as follows. Given a warm/cold process with its system and cost parameters and a demand stream; we consider another production process with the same parameters and demands except that it has inﬁnite capacity per period and cannot be kept warm. That is, the two processes are identical except that R-1 and Q-1 for the latter which corresponds to the classical, uncapacitated setting. Let X ¼ f x1,

x2, . . . , xNgdenote the production plan obtained by the Wagner–

Whitin algorithm for this uncapacitated production process for the given demands. Under Heuristic #9, the production plan X ¼ fx1,x2, . . . ,xNgfor the warm/cold process at hand is obtained

by spreading out the production quantities in X over the periods so as to incur the lowest possible costs by taking into account the

capability of keeping the process warm and the physical capacity level. With this heuristic, a production lot starting at u terminates at period u þ ^n1 after which a positive production quantity is encountered for the ﬁrst time in the Wagner–Whitin solution. The stopping rule can be formally stated as xu þ n40 with

n ¼ minfn : xu þ n40,n 4 1g1.

For the illustrative example, we have the Wagner–Whitin solution given by ff12,0,0g,f6,0g,f12,0g,f10,0g,f17,0,0g,f13,0,0gg. Then, the corresponding production plan for the warm/cold process for periods 1 through 15 is ff6,6,0g,f6,0g,f7,5g, f6,4g, f8,5,4g,f8,5,0gg with a corresponding total cost of 71.45.

Note that costs are computed by considering warming effects for consecutive lots even when the stopping rules generate separate lots (beginning with cold setups); that is, heuristics are used to obtain only the production plan but not the costs. For example, under Heuristic #4, the second and third lots generated result in a production run that goes on for ﬁve consecutive periods (without a cold setup in between) with the process being kept warm between periods seven and eight.

For ease of reference, we provide a list of the basis of construction, the stopping rule and lot run length for all of our heuristics inTable 1.

5. Numerical study and discussion

In this section, we present and discuss our ﬁndings in a numerical study.

We conducted our numerical study to investigate four aspects: (i) percentage deviation from the optimal cost for each heuristic; (ii) dominance of heuristics among themselves; (iii) impact of the parameter values and demand patterns on performances of the heuristics, and ﬁnally (iv) impact of planning horizon lengths when production plans are made and executed dynamically on a rolling horizon basis.

For our numerical study, we considered a problem horizon of N ¼300 periods. Demands are generated randomly from a normal distribution with mean

m

ð ¼500Þ and standard deviation

s

with

s

=

m

Af0,0:2,0:5,1g. Demand streams generated have been pre-processed: (i) All demand values have been truncated to integers, and negative demands have been replaced by zero demands. (ii) Given the demand stream from (i) D0

¼ fD0

1, . . . ,D0Ng, the ﬁnal

demand stream D ¼ fD1, . . . ,DNgis obtained by spreading out D0

over the problem horizon starting with the last period such that no ﬁnal demand value exceeds the imposed capacity for a parti-cular experiment; that is, DN¼minðR,D0NÞand DNj¼minðR,ðD0Njþ

½PNi ¼ Nj þ 1ðD0iDiÞÞÞ for 1rjrN1. If PNi ¼ 1 ðD0iDiÞ40, the

generated stream is infeasible with the given production capacity level and has to be discarded. In our numerical study, we have encountered no such infeasible streams. We set unit holding cost rate h¼1, unit production cost c ¼0 and vary unit warming cost

o

Af0:55,0:65,0:75,0:85,0:95g. The cold setup cost is selected as a function of the mean demand rate, K ¼ ½J2₌₂

_m

_{where J may be}

viewed as a proxy for the average length of production lot; we have J A f2,5g. The warm process threshold also varies as a

function of the mean demand rate, Q ¼

am

with

a

Af1,1:3,1:5,2,2:3,3:5g. The tightness of capacity is attained by selecting the imposed physical capacity as R ¼ ð

g

K=

o

Þ þQ with

g

Af0:3,0:4,0:5,0:6,0:7,0:8,0:9,1:0g. Note that

g

corresponds to the ratio of keeping the process warm for one period and a cold setup. We considered the problem of obtaining a production plan (i) statically (when demands for the entire problem horizon are known at the beginning of the problem horizon) and (ii) on a rolling horizon basis with given planning horizon lengths (when demands are revealed sequentially). We used 19 different plan-ning horizon lengths PHL (in terms of number of periods);

(7)

PHL A f2,3, . . . ,19,20g. Note that the static solution may be viewed as having a planning horizon equal to the problem horizon.

Overall, our experimental set contains 480 ( ¼5 2 6 8) parameter instances for each of the four levels of demand variance and a given planning horizon length. (Capacity and warm thresh-old values have been truncated if non-integer for consistency with demands.) The average fraction of zero demand values in the generated replication streams is 0%, 2.32% and 15.72% for

s

=

m

¼0:1, 0.5 and 1 respectively. Across all planning horizon lengths, we have 9600 (¼ 480 19 þ480) experiment instances for each of the four levels of demand. For the static case, an experiment instance and a parameter instance coincide by deﬁni-tion. For each particular experiment instance – a particular combination of system parameter values – we have generated 30 demand stream replications for non-zero demand variances yielding a total of 873,600 (¼ 480 (1þ19) 30 3 plus 480 (1þ19) for

s

¼0) problem instances with 43,680 instances for the static case and the rest for the rolling horizon case.

We discuss our ﬁndings for the static and rolling horizon cases separately; we begin with the former.

5.1. Static case

For the static case, we use as the benchmark the optimal solution to problem (P) with the entire problem horizon. The total cost over the horizon under a particular heuristic THCiand the optimal total cost over the horizon THCoptare computed as the average values across the replications for an experiment instance. For computing the total cost under a particular heuristic for a problem instance, we used the algorithm employing Program_-FindProductionPlan ði0

,N0

Þdirectly by setting i0

¼1 and N0

¼N. For each experiment instance, we measure heuristic performances in terms of percentage deviations from the optimal total cost which

is computed for Heuristic i A f#1, . . . ,#9g as follows:

D

i% ¼

THCiTHCopt

THCopt

100

Thus, for each heuristic, we obtain a distribution of percentage deviations for four different demand variance levels based on replication-averages over all 480 experiment instances. Table 2

provides the average and the ﬁve number summary (maximum, third, second and ﬁrst quantiles and the minimum) of the devia-tions. As

s

increases, percentage deviations also increase for all heuristics. All heuristics have left-skewed performance distribu-tions for all demand variance values. The performances of Heuristics #3, #4 and #5 are more sensitive – in that order – to

s

. When we consider the performances in the average and median percentage deviations from the optimal. Heuristic #1 performs best in both measures. The ranking of other heuristics changes with the respect to the variance in demand and the performance measure. Heuristic #9 ranks very low when demand variance is low; but its ranking improves with the increase in

s

. In terms of average percentage deviation, Heuristics #6, and #9 perform closely; similar behavior is observed with respect to their median performance as

s

increases. In Table 3, we tabulate the fraction of problem instances (expressed in % points) in which a particular heuristic strictly dominates another one in a pairwise fashion. For example, Heuristic #1 strictly dominates Heuristic #2 in 79.1% of all 43,680 problem instances, strictly dominates Heuristic #3 in 98.4% and so forth. An instance in which a heuristic dominates another does not necessa-rily imply that the dominating heuristic gives the best heuristic solution for that instance. There is no one heuristic that dominates nor is dominated by all others in all instances but there are clear winners. Heuristic #1 is the best and Heuristic #3 is the worst in this ranking although, in mean and median deviation percentages, it performs well for low demand variances.

Table 1

Summary of heuristics.

Heuristic Construction basis Stopping rule with ^n

#1 Silver–Meal K þ Cu,u þ n n o K þ Cu,u þ n þ 1 n þ 1 ^ n ¼ 1 þ max n :K þ Cu,u þ n n Z K þ Cu,u þ n þ 1 n þ 1

#2 Part period balancing Cu,u þ n4K

^

n ¼ maxfn : Cu,u þ nrKg

#3 Least Unit Cost K þ Cu,u þ n

Pu þ n1 i ¼ u Di oK þ Cu,u þ n þ 1 Pu þ n i ¼ uDi ^ n ¼ 1 þ max n :K þ Cu,u þ n Pu þ n1 i ¼ u Di ZK þ C_P_{u þ n}u,u þ n þ 1 i ¼ uDi ( )

#4 Least Total Cost Cu,u þ n4K

^

n ¼ arg minn A f ~n , ~n þ 1g9KCu,u þ n9,

~

n ¼ maxfi : Cu,u þ irKg

#5 Economic production quantity _{n 4 c}_qn_=D

^ n ¼ bcqn =Dc #6 Groff K þ ~Cu,u þ n n=N o K þ ~Cu,u þ n þ 1 ðnþ 1Þ=N ^ n ¼ 1 þ max n : nðnþ 1ÞDu þ no ½cqn2 D g (

#7 McLauren’s Order Moment _{K þ C}

u,u þ n4 dCCn

^

n ¼ maxfn : K þ Cu,u þ nr dCCng #8 McLauren’s Order Moment 2 Pu þ n1

i ¼ u Di4 cqn ^ n ¼ maxfn :Pu þ n1i ¼ u Dir cqng #9 Wagner–Whitin xu þ n40 ^ n ¼ minfn : ˇxu þ n40, n 4 1g1

(8)

Next, we present results for each heuristic on the average performance of a particular heuristic vis-a-vis the optimal solu-tion in the instances where it is the best heuristic. The entry for row j in Table 4 shows, respectively, the average percentage deviation of heuristic j and the average percentage deviation of all heuristics from the optimal in the instances when the heuristic is the best (or among the best) and the percentage of such instances in parentheses. For example, for

s

=

m

¼0, Heuristic #1 is the best performer (or among the best performers) in 81.7% of the experiment instances with an average deviation from the optimal of 11.3% while the average deviation for all heuristics is 20.8%. If a heuristic has no entries (e.g., Heuristic #4), it implies that the heuristic has never performed best for that category of demand. Heuristics #1 and #9 turn out to be best performers for a large fraction of experiment instances in the overall, followed by Heuristic #3. In Table 4, we also tabulate the percentage of problem instances in which a particular heuristic has been found to be the best heuristic, and the percentage of instances in which it is within 2% and 5% proximity of the best heuristic performance. For example, Heuristic #1 results in the best cost performance in 42.9% of the problem instances, is within 2% of the best cost performance in 80.8% of the instances, within 5% of the best cost performance in 98.2% of the instances. Heuristic #1 ranks ﬁrst and Heuristic #7 ranks last by all performance criteria.

Finally, we attempt to identify the range of experiment parameter values where each particular heuristic is the best or among the best. Identiﬁcation of such operating environment

Table 2

The average and ﬁve number summary (maximum, third, second, ﬁrst quantiles, minimum) of deviations from the benchmark (averaged over 30 replications and 19 planning horizons). The benchmark for the static case is the optimal solution; the benchmark for the rolling horizon case is the rolling horizon DP solution.

s/m¼0 s=m¼0:2 (fors=m¼0 ands=m¼0:2) Static #1 12.6;(50.1;14.8;2.7;0.2;0.0) 16.1;(53.2;24.1;5.4;1.8;1.0) #2 15.9;(83.3;25.2;4.3;0.3;0.0) 19.7;(82.8;32.2;5.9;2.0;1.5) #3 12.7;(50.1;14.8;2.7;0.2;0.0) 18.1;(59.0;29.7;9.0;2.6;1.7) #4 22.0;(96.4;38.1;9.0;0.3;0.0) 26.6;(98.8;43.7;9.1;6.0;4.1) #5 22.8;(134.6;35.1;4.7;0.2;0.0) 27.4; 132.7;42.2;16.1;1.9;1.5) #6 22.6;(134.6;35.1;4.5;0.2;0.0) 26.7;(133.9;43.1;9.5;3.3;1.3) #7 28.2;(135.2;39.9;11.5;9.0;0.0) 31.8;(138.6;43.4;18.6;10.4;5.0) #8 22.7;(134.6;35.1;4.7;0.2;0.0) 26.7;(132.6;42.3;14.5;2.1;1.5) #9 23.5;(134.6;35.1;9.4;0.2;0.0) 26.5;(132.1;40.4;12.8;1.4;0.1) Rolling horizon #1 10.8;(50.0;11.9;0.0;0.0; 8.9) 14.5;(53.3;23.0;3.8;1.0; 2.7) #2 14.2;(83.3;24.6;0.2;0.0;0.0) 18.9;(82.7;31.0;6.9;1.4; 2.3) #3 10.8;(50.0;11.9;0.0;0.0; 8.9) 16.5;(59.2;27.9;7.4;1.7; 1.9) #4 22.7;(96.3;35.0;11.1;5.7; 8.9) 25.6;(98.5;41.6;10.2;5.5;0.2) #5 21.3;(134.7;33.0;4.7;0.0; 8.8) 26.2;131.6;39.2;14.8;1.4; 2.2) #6 21.1;(134.7;33.0;4.4;0.0; 8.9) 25.9;(133.6;40.0;12.6;2.8; 2.8) #7 21.4;(134.7;33.0;4.7;0.0; 8.8) 26.3;(128.0;38.5;14.4;3.7; 2.1) #8 21.3;(134.7;33.0;2.2;0.0; 8.8) 25.1;(132.3;39.7;11.9;1.6; 2.3) #9 20.7;(134.7;35.2;1.4;0.0;0.0) 25.2;(131.7;38.1;12.1;1.2;0.0) (fors=m¼0:5 ands=m¼1) s=m¼0:5 s=m¼1 Static #1 19.0;(56.8;29.0;8.4;5.5;3.0) 21.4;(54.5;36.7;11.2;8.1;4.1) #2 23.7;(81.1;38.2;10.6;7.8;3.9) 26.1;(71.2;42.0;14.7;10.6;5.5) #3 29.6;(67.4;38.0;27.3;15.4;12.3) 40.7;72.1;47.9;39.2;33.6;21.5) #4 33.4;(104.2;49.2;18.7;14.8;11.1) 48.0;107.7;56.8;39.8;32.0;29.4) #5 36.3;(134.8;51.8;25.1;11.4;9.9) 43.0;121.9;51.6;33.0;24.6;21.6) #6 29.3;(131.0;43.8;11.4;6.4;2.8) 29.0;(111.5;43.5;12.4;7.7;4.2) #7 38.3;(139.1;55.1;22.0;15.6;11.8) 42.2;131.2;52.0;29.6;23.2;16.7) #8 31.2;(132.6;43.3;19.9;7.8;3.9) 30.7;(119.8;45.6;16.3;8.8;4.9) #9 28.7;(132.3;43.3;11.8;5.4;0.2) 28.5;(117.6;43.3;12.2;6.8;1.6) Rolling horizon #1 17.6;(56.7;28.9;6.9;4.0;0.5) 20.0;(54.4;36.2;9.9;6.5;2.1) #2 23.0;(81.0;35.8;11.4;7.6;1.7) 24.7;(71.3;41.4;13.1;9.0;4.3) #3 27.5;(67.5;35.0;24.0;12.9;7.3) 37.1;(62.1;45.0;35.7;29.6;16.5) #4 30.7;(102.1;45.9;16.7;12.4;6.6) 34.8;94.3;47.3;23.5;19.0;13.7) #5 37.2;(130.3;55.4;24.6;11.4;7.0) 60.4;398.6;66.3;39.2;25.8;17.0) #6 30.1;(132.4;45.7;16.8;6.6;0.7) 34.7;(189.3;52.2;24.4;9.6;3.8) #7 35.8;(134.4;53.2;22.6;11.3;5.1) 50.4;393.9;58.5;33.0;16.8;10.6) #8 30.2;(132.3;43.7;17.9;7.2;1.7) 33.9;(119.7;51.9;25.1;9.5;4.1) #9 27.8;(131.8;42.9;12.4;5.3;0.0) 27.9;(117.6;42.0;12.6;6.6;1.2) Table 3

Fraction of problem instances (in % points) of dominance in pairwise comparison.

Heuristic #1 #2 #3 #4 #5 #6 #7 #8 #9 Static #1 – 79.1 98.4 84.6 59.9 92.3 81.6 96.0 81.4 #2 20.5 – 93.1 59.8 33.1 76.6 59.0 81.4 47.8 #3 1.4 6.7 – 11.3 8.3 41.6 41.0 49.4 9.1 #4 14.9 39.2 88.6 – 24.7 79.3 58.4 78.0 37.5 #5 39.6 65.9 91.7 74.3 – 83.8 63.7 80.1 59.6 #6 7.1 22.4 58.4 17.8 15.2 – 52.2 68.1 11.7 #7 12.9 40.5 58.9 40.9 35.9 45.8 – 61.2 32.1 #8 3.6 18.1 50.4 21.6 19.5 31.2 38.3 – 8.2 #9 15.4 51.8 90.8 54.8 39.9 84.5 62.6 90.3 – Rolling horizon #1 – 87.5 80.5 97.7 94.5 92.2 96.9 93.1 72.0 #2 9.3 – 60.7 87.0 84.7 59.6 87.6 56.9 44.5 #3 13.3 33.9 – 54.1 53.6 45.9 55.9 43.6 40.2 #4 2.1 11.7 45.7 – 63.2 28.9 46.4 29.5 25.8 #5 4.9 11.2 44.2 36.2 – 25.2 31.3 21.2 20.3 #6 7.2 39.9 53.6 70.7 73.8 – 67.7 55.2 32.9 #7 2.6 11.8 43.6 53.3 67.3 31.3 – 28.9 27.8 #8 6.3 31.0 55.6 70.2 76.6 43.8 70.0 – 28.2 #9 27.5 54.9 59.3 74.0 79.0 66.5 71.6 71.2 –

(9)

characteristics would help practitioners choose a priori the best heuristic for a given operational setting. InFigs. 1–4, we illustrate (as blackened) the regions of the experiment set for which a particular heuristic is the best (or among the best). Based on this illustration, we can make the following observations. For

s

¼0, for J¼2, we see that all heuristics except #7 show comparable performance. When J¼5, however, a delineation among the heuristics starts to emerge. Heuristics #1, #2 and #3 dominate the rest but, we see that there are regions for all heuristics in which they are the best (or among the best). (See Fig. 1.) As demand variability increases ð

s

=

m

¼0:2Þ, the deviation increases further. For low J, Heuristics #1 and #9 almost complement each other; that is, one is the best in the regions where the other is not. For high values of

a

, Heuristic #9 dominates. We see fewer instances where Heuristics #2, #4, #5, and #6 are the best (or among the best) performers. Heuristics #3, #7 and #8 are never among the best performers in this region. For high J, Heuristic #1 is the best for the large majority of experiment instances (especially for low

a

) followed by Heuristics #2, #3 and #4. For low and high ends of

a

values, Heuristic #1 dominates the rest of the heuristics. There are also instances where Heuristics #6, #7 and #8 are among the best performers; but Heuristics #5 and #9

are never the best in this region. (SeeFig. 2.) For

s

=

m

¼0:5 and low J, Heuristics #1 and #9 are the best for most of the instances followed by only Heuristic #6; the rest is never among the best performers. We observe the complementary behavior of Heuristics #1 and #9 in this case as well. For low

a

, Heuristic #1 is the best whereas Heuristic #9 dominates for high

a

values. For high J, Heuristic #1 dominates clearly except for moderate values of

a

and high values of

g

. It is complemented in those regions by Heuristic #3. Heuristic #9 starts to show presence for large

a

and

g

. (See Fig. 3.) For high demand variance (

s

=

m

¼1), Heuristic #9 dominates for low J and high

a

values at the expense of Heuristic #1. Heuristic #6 also emerges as the best heuristic complementing Heuristic #9. For high J and low

a

, Heuristic #1 is the sole best heuristic but Heuristics #3 and #9 dominate it in the rest of the experiment instances. For moderate

a

, Heuristic #3 is more likely to be the better than Heuristic #9. (SeeFig. 4.)

Next, we discuss our ﬁndings for the rolling horizon case. 5.2. Rolling horizon case

The rolling horizon case corresponds to the problem of obtaining the production plan dynamically as time progresses

5 1 1.3 1.5 2 2.3 3.5

Fig. 1. Illustration of the experiment instances for which a heuristic is the best performer (or among the best),s=m¼0. Table 4

Mean performances of individual heuristics and all heuristics in the experiment instances in which a particular heuristic is best or among the best and fraction of problem instances (in % points) in which a particular heuristic is the best (or among the best) or Whitin 2% and 5% of the best heuristic solution for the static case. (n

corresponds to 3 instances out of 14,400.) s=m¼0 s=m¼0:2 s=m¼0:5 s=m¼1 Overall r2% r5% #1 11.3/20.8 (81.7) 24.0/38.3 (47.8) 26.3/40.5 (48.4) 31.2/47.0 (31.2) 26.3/40.9 (42.9) (80.8) (98.2) #2 8.4/16.8 (66.7) 22.9/42.6 (14.5) 16.6/29.8 (3.4) 18.2/30.3 (0.7) 20.2/37.3 (6.9) (31.5) (70.0) #3 10.9/20.1 (82.3) 28.5/48.9 (20.6) 30.6/49.9 (10.0) 35.8/57.3 (15.3) 30.2/50.1 (16.0) (35.4) (49.0) #4 3.8/7.2 (45.6) 15.4/21.5 (4.6) –/– –/– 12.5/17.9 (2.0) (5.3) (15.7) #5 0.5/2.4 (46.7) 6.6/9.4 (2.8) –/– –/– 4.4/6.9 (1.4) (14.1) (23.4) #6 0.4/2.4 (47.7) 5.8/9.7 (4.4) 7.0/16.3(5.6) 9.4/24.9 (6.7) 7.0/16.6 (6.0) (34.3) (67.3) #7 6.0/10.5 (14.2) 39.6/43.4 (1.0) 34.5/36.6 (0.0n ) –/– 28.7/32.7 (0.5) (2.2) (12.3) #8 0.5/2.4 (46.9) 16.0/18.7 (2.1) 21.7/26.2 (0.3) 8.8/22.1 (1.5) 9.9/ 15.4 (1.8) (25.5) (61.2) #9 0.6/2.1 (43.8) 0.9/3.8 (27.9) 4.7/10.6 (32.6) 8.4/19.6 (44.6) 5.2/12.5 (35.1) (53.8) (65.7)

(10)

and demands within a planning horizon are revealed sequentially. (That is, demands become known within a planning horizon as the horizon rolls; therefore, the deterministic demand condition forProposition 1hold for the new horizon length.) For this case, we use as the benchmark the dynamic programming (DP) solu-tion of the problem obtained on a rolling basis with the given planning horizon; that is, we solved problem (P) sequentially as the problem horizon extended over time until we reached period

N as the end of the planning horizon. The total cost over the entire problem obtained thus is denoted by THCDPSand is computed as the average value across 30 replications (and nineteen different planning horizons for a parameter instance when so noted). We note that the benchmark rolling horizon DP solution may not give the best cost; this is illustrated by our results, especially for short planning horizon lengths. (We discuss this ﬁnding later below.) The total cost over the horizon under a particular heuristic THCiis

Fig. 2. Illustration of the experiment instances for which a heuristic is the best performer (or among the best),s=m¼0:2.

(11)

computed also similarly as an average. To compute the total cost under a particular heuristic for a problem instance, we used the algorithm whose pseudo-code is given below.

begin j :¼ 1 ^ k :¼ 1 while ðjrNPHLþ1Þ i0 :¼ ^k N0_{:¼ j þ PHL1}

call Program_FindProductionPlan ði0

,N0

Þ j :¼ j þ 1

^

k :¼ last cold setup such that ^krj end

The heuristic performance for each parameter instance for Heuristic iA f#1, . . . ,#9g is found as follows:

D

i% ¼

THCiTHCDPS

THCDPS

100

The analysis and discussion of our ﬁndings follow closely that for the static case above.

Table 2 provides the average and the ﬁve number summary (maximum, third, second and ﬁrst quantiles and the minimum) of the deviations. As

s

increases, percentage deviations also increase for all heuristics. All heuristics have left-skewed performance distributions for all demand variance values. The performances of Heuristics #3, #5 and #7 are more sensitive – in that order – to

s

with maximum deviations under Heuristics #5 and #7 increasing while that of Heuristic #3 decreases. Across all variance levels, Heuristic #1 performs best in terms of both the average and the median. The ranking of other heuristics changes with the respect to the variance in demand and the performance measure. For deterministic demand, there are multiple heuristics resulting in the same performance. As demand variance increases, the difference among average percentage deviations also increases. Note that for deterministic demand Heuristics #1 and #2 have zero values for the median percentage deviation.

In Table 3, we tabulate the fraction of problem instances (expressed in % points) in which a particular heuristic dominates another one in a pairwise fashion. Heuristic #1 is the best and Heuristic #5 is the worst in this ranking.

Fig. 4. Illustration of the experiment instances for which a heuristic is the best performer (or among the best),s=m¼1.

Table 5

Mean performances of individual heuristics and all heuristics in the experiment instances in which a particular heuristic is best or among the best and fraction of problem instances (in % points) in which a particular heuristic is the best (or among the best) or Whitin 2% and 5% of the best heuristic solution for the rolling horizon case.

s=m¼0 s=m¼0:2 s=m¼0:5 s=m¼1 Overall r2% r5% #1 9.5/18.6 (93.8) 19.1/32.4 (55.3) 20.1/34.8 (67.1) 23.1/43.6 (51.5) 16.8/58.0 (66.9) (92.1) (99.4) #2 7.4/16.2 (72.1) 21.6/46.4 (14.1) –/– –/– 9.7/0.0 (21.5) (31.4) (68.5) #3 9.5/19.9 (93.8) 19.2/39.7 (19.8) 29.9/51.4 (10.4) 31.4/53.2 (16.3) 14.9/87.7 (35.1) (37.1) (50.7) #4 5.9/10.6 (27.1) 13.0/15.6 (3.1) –/– –/– 6.6/0.0 (7.6) (3.9) (17.6) #5 0.8/2.3 (49.9) 0.8/2.4 (1.6) –/– –/– 0.8/0.0 (12.9) (14.4) (23.4) #6 0.5/2.3 (51.0) 20.0/22.1 (2.4) –/– 9.1/15.3(0.2) 1.4/1.0 (13.4) (26.7) (58.7) #7 0.8/2.0 (47.9) 11.7/13.6 (0.1) –/– –/– 0.8/ 0.0 (12.0) (8.8) (27.0) #8 0.8/2.3 (49.9) 25.8/28.3 (2.1) –/– –/– 1.8/ 0.0 (13.0) (22.0) (53.1) #9 1.1/3.2 (46.0) 0.6/2.7 (21.5) 2.8/8.9 (22.5) 7.5/25.0 32.1) 3.0/ 4.2 (30.5) (47.1) (62.1)

(12)

Next, we present results for each heuristic on the average performance of a particular heuristic vis a vis the optimal solution in the instances where it is the best heuristic. The entry for column j inTable 5shows, respectively, the average percentage deviation of heuristic j and the average percentage deviation of all heuristics from the optimal in the instances when the heuristic is the best (or among the best) and the percentage of such instances in parentheses. Heuristic #1 is the best performer overall when we compare the fraction of dominated instances and the deviation percentages. In Table 5, we also tabulate the percentage of problem instances in which a particular heuristic has been found to be the best heuristic, and the percentage of instances in which it is within 2% and 5% proximity of the best heuristic performance. Heuristic #1 ranks ﬁrst and Heuristics #4, #5 and #7 are worst performers by all criteria.

Next, as in the static case, we provide, in Figs. 1–4, an illustration (as blackened) of the regions of the parameter set for which a particular heuristic is the best (or among the best) for one short and one longer planning horizon (PHL¼3 and 18). We discuss the short planning horizon case ﬁrst. For

s

¼0, for J ¼2, we see that all heuristics show comparable performance except Heuristics #4 for moderate to large

a

. When J¼ 5, however, a delineation among the heuristics starts to emerge. Heuristics #1, #2 and #3 dominate the rest but, we see that there are regions for all heuristics in which they are the best (or among the best). (See

Fig. 1.) As demand variability increases ð

s

=

m

¼0:2Þ in Fig. 2, certain heuristics start to emerge as the sole best. Speciﬁcally, Heuristic #1 is the best performer over a vast majority of parameter instances. For large J and

a

, Heuristics #2 and especially #3 become better. Heuristics #7 and #9 are never the best; the rest perform well sporadically. For

s

=

m

¼0:5, (inFig. 3) Heuristic #1 is clearly the best performer with Heuristic #3 for some instances with large J and moderate

a

, and Heuristic #9 only for low J and large

a

. The rest of the heuristics is never the best. As demand variance increases further ð

s

=

m

¼1Þ inFig. 4, Heuristic #1 deteriorates and Heuristics #9 and #3 replace it. Heuristic #6 performs well for few instances of large J and moderate

a

. For longer planning horizons, there is not a discernable difference compared to short horizon for

s

¼0; for large J, Heuristic #9 deteriorates for large

a

but improves slightly for moderate

a

.

For moderate

a

and large J, Heuristics #5, #6, and #8 start to perform well. As demand variance increases, Heuristics #1, #3 and #9 start to dominate the rest with Heuristic #3 being best for large J and moderate

a

, and #9 for small J and moderate to high

a

.

Next, we present our observations on the impact of planning horizon length on heuristic performance. In terms of best per-forming heuristics, there is a clear separation among Heuristics #1 and #3 and the rest. The performances deteriorate but stabilize as planning horizon extends. This is to be expected: The benchmark solution approaches the optimal static solution for longer horizon lengths but, as the planning horizon gets very short, it becomes more suboptimal as evidenced by better solutions obtained via Heuristic #1 for such instances.

For

s

=

m

¼0 and low J, Heuristics #1 and #3 are the best performers over all planning horizon lengths, followed by Heuristics #6, #5 and #8. Heuristic #4 performs distinctly badly whereas the remaining heuristics form a cluster with higher deviations. For large J, the performance deteriorates for all heuristics. Heuristics #1 and #3 are again the best performers followed by Heuristics #2, #6 and then by Heuristics #4 and #9. The rest form a separate grouping with very close performances. Overall, the heuristic performance is similar to that for large J since the deviations are larger for large J. For

s

=

m

¼0:2 and low J, Heuristic #1 is clearly the best performer over all planning horizon lengths, followed by a grouping comprising Heuristics #8, #2 and #3. Note that for certain short planning horizons, the deviations of Heuristic #1 from the rolling horizon DP solution are negative which indicates that, on average, Heuristic #1 provides better solutions. (We elaborate on this aspect later.) The second group-ing is followed by Heuristics #9 and #6. Heuristic #9 improves over longer planning horizon lengths whereas Heuristic #6 dete-riorates. The rest follow the trend of the second grouping but with larger deviations. For large J, the performance deteriorates for all heuristics. Heuristic #1 is again the best performer followed closely by Heuristic #3, and then by Heuristics #2 and #6. The rest is in a distinctly separate grouping. Overall, Heuristic #1 performs the best followed by Heuristics #3 and #2. The remain-ing heuristics form a distinct groupremain-ing with similar behavior. For

s

=

m

¼0:5 and low J, Heuristic #1 is the best performer over all planning horizon lengths, followed by Heuristics #9 and #2.